[QE-users] Current state of parallelization methods (in-QE vs. library OMP/ELPA)

[Felix Mayr]

Dear all,

after I found a nice speed-up with MKL (I know, that's common knowledge 
- though getting it to pick up Scalapack with OpenMPI isn't so much), 
I'd like to finally know how to build QE to achieve best performance 
(which varies for specific systems/models of course).
Basically I intend to build different permutations with spack, then run 
benchmark systems with it.
I've run into two problems:
- is there still a flag for disabling threaded versions of MKL? for a 
threaded MKL spack always outputs all the link flags, and QE picks up 
the threaded version if it's enabled. documentation still mentions a 
"--with-openmp"-flag, which could mitigate this. That's not a big issue 
however, the question is just, how this works together with the 
--enable-openmp-flag?
- which version of ELPA should I use. Just doing a spack install tried 
to use elpa-2019.(5?) and QE 6.5 which resulted in a missing elpa1.mod. 
Internet search reveals that QE is by default only compatible with older 
elpa and there is a config switch -D __ELPA2018 or similar to use newer 
ones? What should I choose here?

Best Regards


-- 
Felix Mayr, M.Sc.
research assistant

Technical University of Munich
Department of Electrical and Computer Engineering
Simulation of Nanosystems for Energy Conversion

Arcisstraße 21
80333 Munich, Germany

email: felix.mayr at tum.de
phone: +49-89-289-26933